Independent open-source engineering project by Hill Patel.
The problem
An agent that loses its process or context needs to recover its task state and check what already happened before continuing. Blind retries can repeat an external action.
What I built
Cairn provides a typed continuation state, checkpoint and recovery primitives, an optional agent loop, and an effect ledger. Its recovery protocol re-observes the environment and decides whether to retry, skip a completed effect, or escalate an uncertain outcome.
Design and handover
The library uses injected model and environment interfaces so adopters can keep their own models and tools. The repository includes an offline recovery example, API documentation, tests and a reproduction guide.
The documented recovery contracts and recorded experiments define the evidence boundaries. Integrations with a new model or external service require their own validation. This case study describes independent engineering work; customer adoption and business ROI are not measured here.
Independent open-source Python library for agent crash recovery, durable checkpoints and reconciliation of external effects, with reproducible examples and tests.